EEBIC Group

GP Music Composition Project Page

by Brad Johanson, Supervised by Riccardo Poli

Project Overview

My Spring semester mini-project here at the University of Birmingham is on the use of Genetic Programming techniques to allow computers to compose music. Genetic Programming is an Artificial Intelligence technique that evolves "fit" individual programs from an initially random population of programs. In the case of music, fitness can be defined as how pleasing it is to listen to a particular sequence. For now, all of the pieces are short melodic sequences, generated by programs which are a series of commands such as play these two note sequences, or play the note sequence shifted up. The problem with genetic programming is that tens to hundreds of individuals need to be evaluated over many generations before a satisfying individual is found. Since fitness here is measured as how pleasing the sequence sounds, somebody must sit down and listen to potentially thousands of pieces, many of which may be terrible. The long term goal of this project, therefore, is to take data of people rating pieces in GP runs with fewer individuals and generations, and use that to construct a neural net or mathematical function which can rate individuals in the absence of a human. This stand-in can then be used in more extensive runs to generate pleasant sounding musical pieces.

The Interface

The program is run under just about any UNIX/X-Windows environment, with the development system being a 133MHz Pentium system with 32 MB of RAM running under Linux. When run, the simple X-interface shown below comes up and prompts the user to rate each individual in the generation. After this a new generation is created and the process repeats.

Initial Results

This section contains some of the musical sequences generated during the development of the GP system to date. Unless otherwise specified, the piece is the best of run individual. Each piece is accompanied by the details of the parameters and function/terminal sets in use when the sequences were created. Earlier pieces were generated earlier in the development of the software. The GP code is based on the Lil-GP programming system developed by Doug Zongker at Michigan State University.

Future Work

• Comparative Fitness- currently fitness is determined on a scale of 1-100 on a per individual basis. Fitness by rating one individual against another in a round robin competition to find the best will also be tried
• Sequence Length Limiting- current sequences can be very long and plodding. Methods need to be tried to limit the length of sequences, either by tree node limitation, tree depth limitation, or sequence truncation in elements of the function set.
• Sub-Sequence Fitness Evaluation- many of the sequences have interesting parts and boring parts. It would be nice to be able to keep nice sequence sub-trees or ADFs in the population with a higher probability.
• Data Collection- Once the music GP system has been stabilised, data needs to be taken over many runs with different users which can be used to find patterns in "nice sounding" sequences, and for use in creating an automatic evaluating agent.
• Sequence Evaluator- collected data needs to be used to create an evaluating agent to stand in for humans in larger runs.

Program Source

The following gzipped tar file contains the source code for the GP-Music System. With minor modifications it should run on any Unix workstations with audio output. The xforms library needs to be installed on your workstation, and a means of playing XM mod files to the audio on your workstation is needed. A README file is included in the tar file with more details.

gp-music0.9.tar.gz